[Mesa-dev] [PATCH 4/6] Add texstore functions for BPTC-compressed textures
Neil Roberts
neil at linux.intel.com
Tue Jul 22 12:09:59 PDT 2014
This adds compressors for all four of the BPTC compressed-texture formats. For
the RGB and SRGB normalized BPTC textures it works by first compressing each
4x4 block using the existing DXT3 compressor and then converting it to a BPTC
block. The BPTC block loses one bit of information on the green component of
the endpoints and one bit from the alpha values.
The compressor for the two half-float formats is written from scratch and
takes a very simple approach. It always uses mode 3 of the BPTC format and
picks the two endpoints by dividing the texels into those which have more or
less than the average luminance of the block and then calculating an average
color of the texels within each division.
The first approach using DXT3 gives much better results but neither is really
very good. However it's probably not really sensible to try to use BPTC
compression at runtime because for example with the Nvidia compression tool it
can take in the order of an hour to compress a full-screen image. With that in
mind I don't think it's worth having a proper compressor in Mesa and this
approach gives reasonable results for a usage that is basically a corner case.
---
src/mesa/main/texcompress_bptc.c | 533 +++++++++++++++++++++++++++++++++++++++
src/mesa/main/texcompress_bptc.h | 10 +
src/mesa/main/texstore.c | 10 +
3 files changed, 553 insertions(+)
diff --git a/src/mesa/main/texcompress_bptc.c b/src/mesa/main/texcompress_bptc.c
index 0068975..48271793 100644
--- a/src/mesa/main/texcompress_bptc.c
+++ b/src/mesa/main/texcompress_bptc.c
@@ -29,6 +29,7 @@
#include <stdbool.h>
#include "texcompress.h"
#include "texcompress_bptc.h"
+#include "texcompress_s3tc.h"
#include "texstore.h"
#include "macros.h"
#include "format_unpack.h"
@@ -69,6 +70,12 @@ struct bptc_float_mode {
struct bptc_float_bitfield bitfields[24];
};
+struct bit_writer {
+ uint8_t buf;
+ int pos;
+ uint8_t *dst;
+};
+
static const struct bptc_unorm_mode
bptc_unorm_modes[] = {
/* 0 */ { 3, 4, false, false, 4, 0, true, false, 3, 0 },
@@ -956,3 +963,529 @@ _mesa_get_bptc_fetch_func(mesa_format format)
return NULL;
}
}
+
+static int
+get_alpha_index(int min, int max, int value)
+{
+ if (min == max)
+ return 0;
+ else
+ return (value - min) * 7 / (max - min);
+}
+
+static void
+extract_rgb(uint16_t src, uint8_t *dst)
+{
+ dst[0] = src >> 11;
+ /* We need to lose one bit of the green component */
+ dst[1] = ((src >> 6) & 0x1f);
+ dst[2] = src & 0x1f;
+}
+
+static void
+write_bits(struct bit_writer *writer, int n_bits, int value)
+{
+ do {
+ if (n_bits + writer->pos >= 8) {
+ *(writer->dst++) = writer->buf | (value << writer->pos);
+ writer->buf = 0;
+ value >>= (8 - writer->pos);
+ n_bits -= (8 - writer->pos);
+ writer->pos = 0;
+ } else {
+ writer->buf |= value << writer->pos;
+ writer->pos += n_bits;
+ break;
+ }
+ } while (n_bits > 0);
+}
+
+static void
+convert_dxt3_to_mode4(const uint8_t *src,
+ uint8_t *dst)
+{
+ uint8_t rgb[2][4];
+ uint8_t t[4];
+ uint8_t alpha0, alpha1;
+ bool invert_rgb_indices;
+ bool invert_alpha_indices;
+ int min_alpha = 0xf, max_alpha = 0x0, alpha;
+ struct bit_writer writer;
+ int endpoint, component;
+ int i, value;
+
+ extract_rgb(src[8] | (src[9] << 8), rgb[0]);
+ extract_rgb(src[10] | (src[11] << 8), rgb[1]);
+
+ /* In DXT3 the order of the endpoints is irrelevant. However in BPTC the
+ * order needs to be set so that the anchor index (in this case index 0)
+ * has the most-significant bit set to zero. Therefore if the high bit of
+ * the first index is set then we need to swap the order. Because the order
+ * of the DXT indices is different, the most-significant bit effectively
+ * comes from bit 0 */
+ if (src[12] & 1) {
+ memcpy(t, rgb[0], sizeof t);
+ memcpy(rgb[0], rgb[1], sizeof t);
+ memcpy(rgb[1], t, sizeof t);
+ invert_rgb_indices = true;
+ } else {
+ invert_rgb_indices = false;
+ }
+
+ /* We need to truncate the 4 alpha bits to 3 so in order to get the best
+ * use of the bits we can calculate the range of alpha values used and set
+ * the endpoints to that */
+ for (i = 0; i < BLOCK_SIZE * BLOCK_SIZE / 2; i++) {
+ alpha = src[i] & 0x0f;
+ if (alpha < min_alpha)
+ min_alpha = alpha;
+ if (alpha > max_alpha)
+ max_alpha = alpha;
+ alpha = src[i] >> 4;
+ if (alpha < min_alpha)
+ min_alpha = alpha;
+ if (alpha > max_alpha)
+ max_alpha = alpha;
+ }
+
+ /* We may also need to swap the alpha values to make the anchor index have
+ * a zero most-significant bit */
+ if (get_alpha_index(min_alpha, max_alpha, src[0] & 0x0f) & 0x04) {
+ alpha0 = max_alpha;
+ alpha1 = min_alpha;
+ invert_alpha_indices = true;
+ } else {
+ alpha0 = min_alpha;
+ alpha1 = max_alpha;
+ invert_alpha_indices = false;
+ }
+
+ /* Extend the alpha endpoints to 6 bits */
+ alpha0 <<= 2;
+ if (alpha0 & 4)
+ alpha0 |= 3;
+ alpha1 <<= 2;
+ if (alpha1 & 4)
+ alpha1 |= 3;
+
+ writer.pos = 0;
+ writer.buf = 0;
+ writer.dst = dst;
+
+ write_bits(&writer, 5, 0x10); /* mode 4 */
+ write_bits(&writer, 2, 0); /* rotation */
+ write_bits(&writer, 1, 0); /* index selection */
+
+ for (component = 0; component < 3; component++)
+ for (endpoint = 0; endpoint < 2; endpoint++)
+ write_bits(&writer, 5, rgb[endpoint][component]);
+ write_bits(&writer, 6, alpha0);
+ write_bits(&writer, 6, alpha1);
+
+ /* The first index is special because it has one less bit */
+ write_bits(&writer, 1, (src[12] ^ (src[12] >> 1) ^ invert_rgb_indices) & 1);
+
+ for (i = 1; i < BLOCK_SIZE * BLOCK_SIZE; i++) {
+ value = (src[i / 4 + 12] >> ((i % 4) * 2)) & 3;
+ /* The order of the RGB indices in DXT is different */
+ value = ((value & 1) << 1) | ((value & 1) ^ (value >> 1));
+ if (invert_rgb_indices)
+ value = 3 - value;
+ write_bits(&writer, 2, value);
+ }
+
+ /* The first index is special because it has one less bit */
+ value = get_alpha_index(min_alpha, max_alpha, src[0] & 0x0f);
+ if (invert_alpha_indices)
+ value = 7 - value;
+ write_bits(&writer, 2, value);
+
+ for (i = 1; i < BLOCK_SIZE * BLOCK_SIZE; i++) {
+ value = get_alpha_index(min_alpha, max_alpha,
+ (src[i / 2] >> ((i % 2) * 4)) & 0xf);
+ if (invert_alpha_indices)
+ value = 7 - value;
+ write_bits(&writer, 3, value);
+ }
+}
+
+GLboolean
+_mesa_texstore_bptc_rgba_unorm(TEXSTORE_PARAMS)
+{
+ struct gl_pixelstore_attrib temp_packing;
+ uint8_t block[BLOCK_BYTES];
+ uint8_t *blockPtr = block;
+ GLboolean res;
+ int dst_row_diff;
+ uint8_t *dst = dstSlices[0];
+ int x, y;
+
+ /* It's probably not really sensible to try and use BPTC compression on the
+ * fly because to properly search the entire space takes in the order of an
+ * hour for a full-screen image. Instead we'll just use DXT3 compression
+ * and convert each block to a BPTC block in mode 4. This will lose one bit
+ * of information from the green channel of the endpoint colors and the
+ * alpha will be 3-bit instead of 4.
+ */
+
+ ASSERT(dstFormat == MESA_FORMAT_BPTC_RGBA_UNORM ||
+ dstFormat == MESA_FORMAT_BPTC_SRGB_ALPHA_UNORM);
+
+ temp_packing = *srcPacking;
+ if (temp_packing.RowLength == 0)
+ temp_packing.RowLength = srcWidth;
+
+ dst_row_diff = (dstRowStride >= (srcWidth * 4) ?
+ dstRowStride - (((srcWidth + 3) & ~3) * 4) : 0);
+
+ for (y = 0; y < srcHeight; y += BLOCK_SIZE) {
+ for (x = 0; x < srcWidth; x += BLOCK_SIZE) {
+ temp_packing.SkipPixels = srcPacking->SkipPixels + x;
+ temp_packing.SkipRows = srcPacking->SkipRows + y;
+
+ res = _mesa_texstore_rgba_dxt3(ctx, dims, baseInternalFormat,
+ MESA_FORMAT_RGBA_DXT3,
+ BLOCK_SIZE * 4, /* dstRowstride */
+ &blockPtr,
+ MIN2(srcWidth - x, BLOCK_SIZE),
+ MIN2(srcHeight - y, BLOCK_SIZE),
+ 1, /* srcDepth */
+ srcFormat, srcType, srcAddr,
+ &temp_packing);
+ if (!res)
+ return GL_FALSE;
+
+ convert_dxt3_to_mode4(block, dst);
+ dst += BLOCK_BYTES;
+ }
+
+ dst += dst_row_diff;
+ }
+
+ return GL_TRUE;
+}
+
+static float
+get_average_luminance(int width, int height,
+ const float *src, int src_rowstride)
+{
+ float luminance_sum = 0;
+ int y, x;
+
+ for (y = 0; y < height; y++) {
+ for (x = 0; x < width; x++) {
+ luminance_sum += src[0] + src[1] + src[2];
+ src += 3;
+ }
+ src += (src_rowstride - width * 3 * sizeof (float)) / sizeof (float);
+ }
+
+ return luminance_sum / (width * height);
+}
+
+static float
+clamp_value(float value, bool is_signed)
+{
+ if (value > 65504.0f)
+ return 65504.0f;
+
+ if (is_signed) {
+ if (value < -65504.0f)
+ return -65504.0f;
+ else
+ return value;
+ }
+
+ if (value < 0.0f)
+ return 0.0f;
+
+ return value;
+}
+
+static void
+get_endpoints(int width, int height,
+ const float *src, int src_rowstride,
+ float average_luminance, float endpoints[][3],
+ bool is_signed)
+{
+ float endpoint_luminances[2];
+ float midpoint;
+ float sums[2][3];
+ int endpoint, component;
+ float luminance;
+ float temp[3];
+ const float *p = src;
+ int left_endpoint_count = 0;
+ int y, x, i;
+
+ memset(sums, 0, sizeof sums);
+
+ for (y = 0; y < height; y++) {
+ for (x = 0; x < width; x++) {
+ luminance = p[0] + p[1] + p[2];
+ if (luminance < average_luminance) {
+ endpoint = 0;
+ left_endpoint_count++;
+ } else {
+ endpoint = 1;
+ }
+ for (i = 0; i < 3; i++)
+ sums[endpoint][i] += p[i];
+
+ p += 3;
+ }
+
+ p += (src_rowstride - width * 3 * sizeof (float)) / sizeof (float);
+ }
+
+ if (left_endpoint_count == 0 ||
+ left_endpoint_count == width * height) {
+ for (i = 0; i < 3; i++)
+ endpoints[0][i] = endpoints[1][i] =
+ (sums[0][i] + sums[1][i]) / (width * height);
+ } else {
+ for (i = 0; i < 3; i++) {
+ endpoints[0][i] = sums[0][i] / left_endpoint_count;
+ endpoints[1][i] = sums[1][i] / (width * height - left_endpoint_count);
+ }
+ }
+
+ /* Clamp the endpoints to the range of a half float and strip out
+ * infinities */
+ for (endpoint = 0; endpoint < 2; endpoint++) {
+ for (component = 0; component < 3; component++) {
+ endpoints[endpoint][component] =
+ clamp_value(endpoints[endpoint][component], is_signed);
+ }
+ }
+
+ /* We may need to swap the endpoints to ensure the most-significant bit of
+ * the first index is zero */
+
+ for (endpoint = 0; endpoint < 2; endpoint++) {
+ endpoint_luminances[endpoint] =
+ endpoints[endpoint][0] +
+ endpoints[endpoint][1] +
+ endpoints[endpoint][2];
+ }
+ midpoint = (endpoint_luminances[0] + endpoint_luminances[1]) / 2.0f;
+
+ if ((src[0] + src[1] + src[2] <= midpoint) !=
+ (endpoint_luminances[0] <= midpoint)) {
+ memcpy(temp, endpoints[0], sizeof temp);
+ memcpy(endpoints[0], endpoints[1], sizeof temp);
+ memcpy(endpoints[1], temp, sizeof temp);
+ }
+}
+
+static void
+write_rgb_indices(struct bit_writer *writer,
+ int src_width, int src_height,
+ const float *src, int src_rowstride,
+ float endpoints[][3])
+{
+ float luminance;
+ float endpoint_luminances[2];
+ int endpoint;
+ int index;
+ int y, x;
+
+ for (endpoint = 0; endpoint < 2; endpoint++) {
+ endpoint_luminances[endpoint] =
+ endpoints[endpoint][0] +
+ endpoints[endpoint][1] +
+ endpoints[endpoint][2];
+ }
+
+ /* If the endpoints have the same luminance then we'll just use index 0 for
+ * all of the texels */
+ if (endpoint_luminances[0] == endpoint_luminances[1]) {
+ write_bits(writer, BLOCK_SIZE * BLOCK_SIZE * 4 - 1, 0);
+ return;
+ }
+
+ for (y = 0; y < src_height; y++) {
+ for (x = 0; x < src_width; x++) {
+ luminance = src[0] + src[1] + src[2];
+
+ index = ((luminance - endpoint_luminances[0]) * 15 /
+ (endpoint_luminances[1] - endpoint_luminances[0]));
+ if (index < 0)
+ index = 0;
+ else if (index > 15)
+ index = 15;
+
+ assert(x != 0 || y != 0 || index < 8);
+
+ write_bits(writer, (x == 0 && y == 0) ? 3 : 4, index);
+
+ src += 3;
+ }
+
+ /* Pad the indices out to the block size */
+ if (src_width < BLOCK_SIZE)
+ write_bits(writer, 4 * (BLOCK_SIZE - src_width), 0);
+
+ src += (src_rowstride - src_width * 3 * sizeof (float)) / sizeof (float);
+ }
+
+ /* Pad the indices out to the block size */
+ if (src_height < BLOCK_SIZE)
+ write_bits(writer, 4 * BLOCK_SIZE * (BLOCK_SIZE - src_height), 0);
+}
+
+static int
+get_endpoint_value(float value, bool is_signed)
+{
+ GLhalfARB half;
+
+ if (is_signed) {
+ half = _mesa_float_to_half(value);
+ return half >> 6;
+ } else {
+ if (value <= 0.0f)
+ return 0;
+
+ half = _mesa_float_to_half(value);
+
+ return half >> 5;
+ }
+}
+
+static void
+compress_rgb_float_block(int src_width, int src_height,
+ const float *src, int src_rowstride,
+ uint8_t *dst,
+ bool is_signed)
+{
+ float average_luminance;
+ float endpoints[2][3];
+ struct bit_writer writer;
+ int component, endpoint;
+ int endpoint_value;
+
+ average_luminance =
+ get_average_luminance(src_width, src_height, src, src_rowstride);
+ get_endpoints(src_width, src_height, src, src_rowstride,
+ average_luminance, endpoints, is_signed);
+
+ writer.dst = dst;
+ writer.pos = 0;
+ writer.buf = 0;
+
+ write_bits(&writer, 5, 3); /* mode 3 */
+
+ /* Write the endpoints */
+ for (endpoint = 0; endpoint < 2; endpoint++) {
+ for (component = 0; component < 3; component++) {
+ endpoint_value =
+ get_endpoint_value(endpoints[endpoint][component], is_signed);
+ write_bits(&writer, 10, endpoint_value);
+ }
+ }
+
+ write_rgb_indices(&writer,
+ src_width, src_height,
+ src, src_rowstride,
+ endpoints);
+}
+
+static void
+compress_rgb_float(int width, int height,
+ const float *src, int src_rowstride,
+ uint8_t *dst, int dst_rowstride,
+ bool is_signed)
+{
+ int dst_row_diff;
+ int y, x;
+
+ if (dst_rowstride >= width * 4)
+ dst_row_diff = dst_rowstride - ((width + 3) & ~3) * 4;
+ else
+ dst_row_diff = 0;
+
+ for (y = 0; y < height; y += BLOCK_SIZE) {
+ for (x = 0; x < width; x += BLOCK_SIZE) {
+ compress_rgb_float_block(MIN2(width - y, BLOCK_SIZE),
+ MIN2(height - y, BLOCK_SIZE),
+ src + x * 3 +
+ y * src_rowstride / sizeof (float),
+ src_rowstride,
+ dst,
+ is_signed);
+ dst += BLOCK_BYTES;
+ }
+ dst += dst_row_diff;
+ }
+}
+
+static GLboolean
+texstore_bptc_rgb_float(TEXSTORE_PARAMS,
+ bool is_signed)
+{
+ const float *pixels;
+ const float *tempImage = NULL;
+ GLenum baseFormat;
+ int rowstride;
+
+ if (srcFormat != GL_RGB ||
+ srcType != GL_FLOAT ||
+ ctx->_ImageTransferState ||
+ srcPacking->SwapBytes) {
+ /* convert image to RGB/float */
+ baseFormat = _mesa_get_format_base_format(dstFormat);
+ tempImage = _mesa_make_temp_float_image(ctx, dims,
+ baseInternalFormat,
+ baseFormat,
+ srcWidth, srcHeight, srcDepth,
+ srcFormat, srcType, srcAddr,
+ srcPacking,
+ ctx->_ImageTransferState);
+ if (!tempImage)
+ return GL_FALSE; /* out of memory */
+
+ pixels = tempImage;
+ rowstride = srcWidth * sizeof(float) * 3;
+ } else {
+ pixels = _mesa_image_address2d(srcPacking, srcAddr, srcWidth, srcHeight,
+ srcFormat, srcType, 0, 0);
+ rowstride = _mesa_image_row_stride(srcPacking, srcWidth,
+ srcFormat, srcType);
+ }
+
+ compress_rgb_float(srcWidth, srcHeight,
+ pixels, rowstride,
+ dstSlices[0], dstRowStride,
+ is_signed);
+
+ free((void *) tempImage);
+
+ return GL_TRUE;
+}
+
+GLboolean
+_mesa_texstore_bptc_rgb_signed_float(TEXSTORE_PARAMS)
+{
+ ASSERT(dstFormat == MESA_FORMAT_BPTC_RGB_SIGNED_FLOAT);
+
+ return texstore_bptc_rgb_float(ctx, dims, baseInternalFormat,
+ dstFormat, dstRowStride, dstSlices,
+ srcWidth, srcHeight, srcDepth,
+ srcFormat, srcType,
+ srcAddr, srcPacking,
+ true /* signed */);
+}
+
+GLboolean
+_mesa_texstore_bptc_rgb_unsigned_float(TEXSTORE_PARAMS)
+{
+ ASSERT(dstFormat == MESA_FORMAT_BPTC_RGB_UNSIGNED_FLOAT);
+
+ return texstore_bptc_rgb_float(ctx, dims, baseInternalFormat,
+ dstFormat, dstRowStride, dstSlices,
+ srcWidth, srcHeight, srcDepth,
+ srcFormat, srcType,
+ srcAddr, srcPacking,
+ false /* unsigned */);
+}
+
diff --git a/src/mesa/main/texcompress_bptc.h b/src/mesa/main/texcompress_bptc.h
index ad5ff50..814548e 100644
--- a/src/mesa/main/texcompress_bptc.h
+++ b/src/mesa/main/texcompress_bptc.h
@@ -27,6 +27,16 @@
#include <inttypes.h>
#include "glheader.h"
#include "texcompress.h"
+#include "texstore.h"
+
+GLboolean
+_mesa_texstore_bptc_rgba_unorm(TEXSTORE_PARAMS);
+
+GLboolean
+_mesa_texstore_bptc_rgb_signed_float(TEXSTORE_PARAMS);
+
+GLboolean
+_mesa_texstore_bptc_rgb_unsigned_float(TEXSTORE_PARAMS);
compressed_fetch_func
_mesa_get_bptc_fetch_func(mesa_format format);
diff --git a/src/mesa/main/texstore.c b/src/mesa/main/texstore.c
index c87e407..bb329df 100644
--- a/src/mesa/main/texstore.c
+++ b/src/mesa/main/texstore.c
@@ -67,6 +67,7 @@
#include "texcompress_rgtc.h"
#include "texcompress_s3tc.h"
#include "texcompress_etc.h"
+#include "texcompress_bptc.h"
#include "teximage.h"
#include "texstore.h"
#include "enums.h"
@@ -3690,6 +3691,15 @@ _mesa_get_texstore_func(mesa_format format)
table[MESA_FORMAT_Z_FLOAT32] = _mesa_texstore_z32;
table[MESA_FORMAT_Z32_FLOAT_S8X24_UINT] = _mesa_texstore_z32f_x24s8;
+ table[MESA_FORMAT_BPTC_RGBA_UNORM] =
+ _mesa_texstore_bptc_rgba_unorm;
+ table[MESA_FORMAT_BPTC_SRGB_ALPHA_UNORM] =
+ _mesa_texstore_bptc_rgba_unorm;
+ table[MESA_FORMAT_BPTC_RGB_SIGNED_FLOAT] =
+ _mesa_texstore_bptc_rgb_signed_float;
+ table[MESA_FORMAT_BPTC_RGB_UNSIGNED_FLOAT] =
+ _mesa_texstore_bptc_rgb_unsigned_float;
+
table[MESA_FORMAT_A_UINT8] = _mesa_texstore_rgba_uint8;
table[MESA_FORMAT_A_UINT16] = _mesa_texstore_rgba_uint16;
table[MESA_FORMAT_A_UINT32] = _mesa_texstore_rgba_uint32;
--
1.9.3
More information about the mesa-dev
mailing list